Garment Factory Surveillance: 60% Reduction in Pilferage Using AI Video Analytics

Fabric pilferage had been a persistent and costly problem for the factory management team for at least three years before this project. The issue manifested in two ways. The first was small-scale daily pilferage: individual workers removing small quantities of fabric — offcuts, end-of-roll remnants, and occasionally full fabric pieces — concealed in clothing or personal bags when leaving the factory at shift end. The second was more significant: coordinated extraction of finished goods from the bonded warehouse, suspected to involve loading dock staff and vehicle crew, with goods removed in shipment vehicles that were not fully manifested.

The factory's finance director conducted a stock reconciliation exercise in 2024 that compared raw material intake records with finished goods output and identified a discrepancy consistent with pilferage losses estimated at BDT 120 lakh per year — approximately USD 100,000. This figure represented roughly 0.6% of annual turnover but was almost entirely profit-margin loss, since the materials themselves had already been purchased. Fabric — particularly the higher-value synthetic fabrics used for export garments — has a ready secondary market in Dhaka's fabric wholesale districts, making it an attractive target.

The factory's existing CCTV system was wholly inadequate for detection or deterrence. The 32-camera DVR system, installed in 2017, had cameras covering only the main production floor aisle ways — there was no coverage of the cutting area where fabric remnants were most accessible, no coverage of the fabric storage area within the warehouse, and no cameras at the loading dock. The DVR recorded over its tapes on a daily cycle. Even when security staff suspected theft, there was no footage to review that would confirm it. The cameras were visible enough to workers that their fixed positions and gaps were well understood.

The BGMEA audit requirement added a compliance dimension to what was already a business problem. BGMEA's security standards for member factories require documented CCTV coverage of production floors, including cutting areas. The 2025 BGMEA annual audit of the factory noted the CCTV gaps on the cutting floor as a compliance deficiency requiring remediation before the next audit cycle.

The factory management had attempted to address pilferage through procedural controls: body checks at factory exit (which generated significant worker relations friction), bag inspection policies (inconsistently enforced), and increased security guard deployment at the loading dock. None of these measures had materially reduced the estimated loss rate. The General Manager acknowledged that procedural controls without supporting technology were ineffective in a facility with 3,500 workers across multiple entry/exit points.

SSAL deployed an 84-camera IP surveillance system with integrated AI video analytics covering all production floors, the bonded warehouse, loading dock, and vehicle gate.

Camera Hardware — Hikvision AcuSense 4MP: 84 Hikvision AcuSense cameras were deployed in three types according to location:

• 60x DS-2CD2143G2-I indoor dome cameras (cutting floor, sewing floor, finishing floor, warehouse interior) — 4MP, H.265+, IR illumination to 40m, IK10 rated.
• 12x DS-2CD2347G2-LU outdoor ColorVu cameras (loading dock, external warehouse gates, factory perimeter) — 4MP with white-light full-color imaging for night identification without IR degradation.
• 12x DS-2CD4585G0-IZS varifocal bullet cameras (vehicle gate, entry/exit points) — 8MP for LPR-capable detail capture at vehicle lanes.

Hikvision DeepinMind NVR — AI Video Analytics: Two Hikvision iDS-9616NXI-I8/16S DeepinMind NVRs provide the AI analytics processing. DeepinMind uses deep learning algorithms running on dedicated neural processing hardware within the NVR itself — analytics processing does not require a separate server. The NVRs support simultaneous analytics processing across all connected camera channels.

The following analytics behaviors were configured:

Zone Intrusion Detection: The fabric storage room in the warehouse (accessible from the production floor via a controlled doorway) was defined as a restricted zone. Any person detected entering this zone outside authorized access hours (07:00-09:00 and 18:00-19:00 only, for fabric issue and receipt) triggers an immediate alert to the security monitoring station. During the commissioning phase, the analytics were tested over two weeks with the threshold calibrated to eliminate false alerts from cleaning staff entering the area.

Line Crossing Detection: Lines were defined at the exit points from the cutting floor (where fabric remnants are most accessible) and at the loading dock perimeter. Any person crossing these lines while carrying an object of significant size (detected by the person-with-object AI classification) triggers an alert. The AI distinguishes between workers carrying factory-approved work materials (trolleys of garments on approved routes) and individuals crossing with unpackaged objects — reducing false alarm rates to near zero.

LPR at Vehicle Gate: The 8MP varifocal cameras at the vehicle gate are configured for Automatic License Plate Recognition. Every vehicle entering and leaving the factory is logged with plate number, timestamp, and camera-captured image. The LPR database is accessible to the security manager and logistics team, enabling cross-reference with shipping manifests and delivery records. Any vehicle without a pre-registered plate number triggers a hold alert to the gate security post.

Storage and Retention: Both DeepinMind NVRs are configured with RAID 1 storage, providing 30-day continuous recording at full resolution across all 84 cameras. Total storage: 2 x 24TB usable capacity. Security staff access the system via Hikvision iVMS-4200 VMS, with the analytics event log providing a searchable alert history in addition to the full video archive.

Security Monitoring Station: A dedicated security monitoring station was set up in the factory security office, with a 4-monitor workstation displaying the AI analytics alert panel on one monitor and live camera grids on the remaining three. The analytics alert panel provides a real-time feed of all triggered events, with one-click access to the relevant camera footage. The security supervisor can acknowledge alerts, add notes, and flag incidents for follow-up investigation.